A mathematical theory for identifying and measuring severity of episodes of care.
OBJECTIVES: We propose and test a method for constructing episodes of care from data within administrative databases and electronic health records. SUBJECTS: We created a measure for severity of episodes of illness for 565 randomly chosen developmentally delayed children who were enrolled in the Medicaid program. DESIGN: Regression analysis was conducted to test the percentage of variance explained by our proposed mathematical model in cost of care. DATA COLLECTION: Data included both hospitalizations and clinic visits obtained from Medicaid programs from one southeastern state. METHODS: For each patient, the likelihood that two diagnoses are part of the same episode is proportional to the similarity of the two diagnoses and to the short time interval between them. When this likelihood exceeds a preset cutoff, then the two diagnoses are part of the same episode. The cutoff is estimated by selecting number of days before two very similar diagnoses are considered to be part of separate episodes. The similarity between two diagnoses is assumed to be proportional to co-occurrence of the two diagnoses within a fixed period (usually 30 days). The severity of an episode was calculated using a Muliplicative Multiattribute Utility model, where severity of each diagnosis is aggregated to estimate the overall severity of the episode. Severity of each diagnosis was assumed to be proportional to average cost of a diagnosis-if patients do not die before care is delivered. The article includes an algorithm that can classify a patient's diagnosis into episodes of care and measure severity of the episodes from date of diagnoses, code for the diagnoses, and charges for the visit. To facilitate integration with existing database, the article includes a Standard Query Language computer program. To evaluate the method of constructing episodes of care, we regressed cost of care on the patient's number of episodes of care within the year, average severity of the episodes within the year, and the interaction between number and average severity of the episodes. RESULTS: The number of episodes (alpha = .001), the average severity of the episodes (alpha = .001), and the product of the two (alpha = .001) had statistically significant relationships to the average cost of the case. The 3 variables together explained 53% of variation in yearly cost of care. CONCLUSIONS: These data suggest that our proposed mathematical approach is reasonable and produces severity scores that are predictive of objective criteria such as cost of care.